Partenaires

CNRS
Logo tutelle
Logo tutelle
Logo tutelle

Rechercher

Sur ce site

Sur le Web du CNRS

Accueil du site > Séminaires > Probabilités Statistiques et réseaux de neurones > Design of Artificial Neural Networks using Evolutionary Computation

Vendredi 29 mars 2002 à 11h00

Design of Artificial Neural Networks using Evolutionary Computation

Francisco Sandoval (Université de Malaga)

Résumé : Artificial Neural Networks (ANNs) offer an attractive paradigm of computation for many applications (pattern recognition, system identification, cognitive modeling, etc.) for a number of reasons including : potential for massively parallel computation, robustness in the presence of noise, resilience to the failure of components, amenability to adaptation and learning, etc. Practical applications of ANNs require the choice of a suitable network topology and the processing functions computed by individual units. However, it is often hard to design good ANNs, because many of the basic principles governing information processing in ANNs are difficult to understand, and the complex interactions among network units usually makes engineering techniques like divide and conquer inapplicable. When complex combinations of behavior approaches are given (such as learning speed, compactness, generalization capacity and resistance to the noise), and the size of the nets grows in dimension and complexity, the approach to its solution by means of the human engineering doesn’t work and it is necessary to appeal to more efficient automated procedures. In this intent of automated solutions it appears the evolutionary techniques, denominated, in a generic way, evolutionary computation. These techniques take form in a group of evolutionary algorithms whose main implementations have been summed up in three approaches, strongly related, but developed in an independent way : genetic algorithms (with links to genetic programming and classifier systems), evolution strategies, and evolutionary programming. All these algorithms respond to a class of population-based stochastic search algorithms, and they have been developed from ideas and principles of natural evolution. An important characteristic of all these algorithms is its search strategy based on the population. The evolutionary algorithms can be used for the design of artificial neural networks, in which, besides the learning, the evolution is another fundamental form of adaptation. This way, the evolutionary algorithms can be used for the realization of diverse tasks, as training of the weights of connection, design of the architecture, adaptation of the learning rules, initialization of the weights, extraction of rules, etc. That is, we try to design artificial neural networks able to adapt to an environment as well as to changes in that environment. The conference will deal with which are the most important characteristics in the evolutionary algorithms, analyzing and comparing their most important constituents, and how these algorithms can be applied in the design of artificial neural networks.

Dans la même rubrique :